CN115359399A - Underground drainage pipeline defect detection and identification method based on improved YOLOX - Google Patents

Abstract

The invention relates to an underground drainage pipeline defect detection and identification method based on improved YOLOX, and provides a method for detecting and identifying defects of an underground drainage pipeline based on the improved YOLOX in order to realize the automation of the detection and identification of the defects of the underground drainage pipeline. Firstly, carrying out frame extraction on an obtained underground drainage pipeline video, carrying out image quality screening and defect classification on the extracted video frame, carrying out image generation on two defect type images of cracks and staggers by adopting a GAN network, selecting a generated image with high image quality, adding the generated image into an underground drainage pipeline defect image data set, and carrying out data set enhancement operation on the obtained underground drainage pipeline defect image data set. And finally, inputting the defect data set of the underground drainage pipeline into an improved YOLOX algorithm to detect and identify the defects of the underground drainage pipeline. The experimental result shows that compared with the manual judgment and the traditional machine learning method, the method can better detect and identify the defects of the underground drainage pipeline.

Description

Underground drainage pipeline defect detection and identification method based on improved YOLOX

Technical Field

The invention relates to an underground drainage pipeline defect detection and identification method based on improved YOLOX, which realizes the automation of underground drainage pipeline defect detection and identification and belongs to the aspect of engineering application.

Background

At present, the most widely applied method for detecting underground drainage pipelines is a Closed Circuit Television (CCTV) visual detection technology, and professionals judge, read and grade the defects of the pipelines through videos and images collected by the CCTV. However, the human interpretation is easily affected by human subjectivity and requires a lot of manpower and cost. The establishment of an intelligent pipeline defect detection and classification system has become an urgent need for the construction and management of urban drainage facilities.

In the conventional defect inspection of the underground drainage pipeline, the defect inspection is mainly carried out through manual visual inspection, however, the method needs to consume a lot of time and manpower, and the result is often subjective due to the influence of human subjectivity. As machine learning advances in various fields, some researchers employ traditional machine learning methods, such as diagnosing defects based on morphological, geometric, and texture features, but the machine learning methods require human manual design of feature extractors, are labor intensive, and are computationally expensive when faced with large data sets. At present, deep learning is widely applied to various image fields, effective features are automatically extracted by using a convolutional neural network based on a deep learning method, and the problems of poor generalization capability, poor robustness and the like of manual features are solved.

The existing image target detection method mainly comprises a single-stage algorithm and a two-stage algorithm, wherein the two-stage algorithm firstly finds out an interested area in an image, namely the position where a target object appears, to obtain a preselected frame, then carries out feature extraction and finally carries out classification, but the two-stage algorithm is slow in speed, and can not meet the requirements for detection tasks with required speed, and the one-stage algorithm directly extracts features in a network to predict object classification and positions, so that the speed is greatly improved. The one-stage algorithm is based on the YOLO series, and mainly comprises the following steps: YOLOV3, YOLOV5, YOLOX, etc., the YOLOX detection algorithm has the best performance in all aspects at present, so it is necessary to design a YOLOX-based underground drainage pipeline defect detection and identification method, which saves human resources and time cost and finds the underground drainage pipeline defects in time.

The YOLOX detection algorithm ingeniously combines excellent progress in the target detection fields of decoupling heads, data enhancement, no anchor points, label classification and the like with YOLO, but the problem of feature aliasing caused by direct fusion of feature maps exists in a feature fusion layer of the YOLOX algorithm, so that the YOLOX structure needs to be improved to detect and identify the defects of the underground drainage pipeline.

Disclosure of Invention

The invention provides an underground drainage pipeline defect detection and identification method based on improved YOLOX, which is used for classifying pipeline defects and intelligently detecting and identifying the defects of underground pipelines.

YOLOX consists of four parts: the device comprises an input layer, a trunk feature extraction layer, a feature fusion layer and a prediction layer. The trunk feature extraction layer is formed by a Focus module and a certain number of stacked CSP modules, a Silu activation function is used, and finally a trunk network feature extraction result is obtained through an SPP module. The feature fusion layer adopts a feature pyramid network FPN structure, the FPN structure conveys strong semantic features from top to bottom, and feature graphs of different scales are fused. And the prediction layer detects the target by adopting three characteristic graphs with different scales. Because the FPN network only simply carries out direct combination when carrying out fusion on the feature graphs, in order to prevent the problem of feature aliasing caused by direct fusion of the feature graphs of different layers, a weight-based feature fusion module is designed, the features of different layers are fused by using weight coefficients, and the feature fusion mode can be expressed as

The method comprises the following steps of fusing two input feature graphs F2 and F1, and enabling the fused features to pass through a replacement attention mechanism module to obtain more detailed information of a target needing attention so as to inhibit other useless information. Then inputting the maximum pooling layer and the average pooling layer respectively, reducing the channels to 1, finally adding the two characteristics and obtaining a weight function by using a sigmoid function.

Meanwhile, information loss mainly occurs in top-level features, and the top-level features in the FPN only contain single-scale features and are not fused with other feature maps, so that the problem is solved by using an improved cavity space pyramid module, a compression excitation module is introduced by using the ASPP thought, multi-scale feature capture is performed on the top-level features, and then richer refined feature maps are obtained through a channel attention mechanism and finally merged.

The method comprises the following specific steps of detecting and identifying the defects of the underground drainage pipeline based on the improved YOLOX:

(1) Firstly, video frame extraction is carried out on the obtained underground drainage pipeline video image according to the frame number at certain intervals.

(2) And screening the extracted video frames according to a certain rule. The screening rule is as follows: for a plurality of images with extremely high similarity, an image quality evaluation algorithm is adopted to evaluate the image quality of the video frames, and underground drainage pipeline images with good quality are screened out to serve as an underground drainage pipeline defect image data set.

(3) And (4) carrying out defect classification on the defect image data set of the underground drainage pipeline, wherein the defect data set is divided into 5 defects of deposition, leakage, tree root invasion, cracks and malting.

(4) Aiming at the problem of too few images with two defect types of cracks and faults, a GAN network is adopted for image generation, two original images are input into a generator G of the GAN network to learn the distribution characteristics of a real image, and a discriminator D compares the generated image with the real image to generate an image with extremely high similarity with the original image.

(5) And screening the images generated through the GAN network, discarding the images with extremely poor effects, further screening the rest images by using an image quality evaluation algorithm, and finally adding the obtained images into the defect image data set of the underground drainage pipeline.

(6) And carrying out data set expansion operation on the data set of the defect image of the underground drainage pipeline, and enhancing the defect image by adopting 5 data enhancement modes of translation, brightness change, noise addition, rotation and mirror image.

(7) Dividing the processed data set of the defect images of the underground drainage pipeline into a training set, a verification set and a test set according to the ratio of 8: 1.

(8) And inputting the underground drainage pipeline defect image data set into a modified YOLOX algorithm for defect detection and identification.

(9) And testing the network obtained by training, counting the test result, and analyzing the accuracy rate, recall rate, average accuracy and average accuracy mean value of the algorithm.

Drawings

FIG. 1 is a flow chart of defect detection and identification of underground drainage pipeline based on improved YOLOX.

FIG. 2 is a block diagram of a weight-based feature fusion module according to the present invention.

FIG. 3 is a schematic diagram of an improved void space pyramid module according to the present invention.

FIG. 4 is a diagram of subsurface pipeline defects generated by a GAN network in accordance with an embodiment of the present invention.

FIG. 5 is a defect diagram of an underground drainage pipeline according to an embodiment of the present invention.

FIG. 6 is a diagram illustrating defect detection and identification of underground drainage pipes according to an embodiment of the present invention.

Detailed Description

The embodiment is as follows:

take underground drainage pipelines in a certain city as an example:

(1) Firstly, acquiring a drainage pipeline defect image from a video, and obtaining an initial underground drainage pipeline defect image through image quality evaluation operation.

(2) Inputting the obtained crack and gap images in the step (1) into a GAN network for image generation, wherein the generated partial images are shown in figure 4, and then all the images are subjected to data enhancement operation to obtain a final underground drainage pipeline defect data set, and the underground pipeline defect partial images are shown in figure 5.

(3) And inputting the defect data set of the underground drainage pipeline into an improved YOLOX algorithm, and outputting a detection identification result through a trunk feature extraction layer, a feature fusion layer and a prediction layer. The results of the detection are shown in FIG. 6.

(4) And (4) counting the detection and identification results, and analyzing the detection and identification accuracy.

Claims (1)

1. An underground drainage pipeline defect detection and identification method based on improved YOLOX is characterized by comprising the following specific processes:

(1) Firstly, extracting video frames of an acquired underground drainage pipeline video image according to a certain interval frame number;

(2) Screening the extracted video frame according to a certain rule, wherein the screening rule is as follows: for a plurality of images with extremely high similarity, carrying out image quality evaluation on video frames by adopting an image quality evaluation algorithm, and screening underground drainage pipeline images with good quality as an underground drainage pipeline defect image data set;

(3) Classifying the defects of the underground drainage pipeline defect image data set, wherein the defects are 5 defects of deposition, leakage, tree root invasion, cracks and malting;

(4) Aiming at the problem of too few images with two defect types of cracks and faults, a GAN network is adopted for image generation, two original images are input into a generator G of the GAN network to learn the distribution characteristics of a real image, and a discriminator D compares the generated image with the real image to generate an image with extremely high similarity to the original image;

(5) Screening the images generated through the GAN network, discarding the images with extremely poor effect, further screening the rest images by using an image quality evaluation algorithm, and finally adding the obtained images into a defect image data set of the underground drainage pipeline;

(6) Carrying out data set expansion operation on the defect image data set of the underground drainage pipeline, and enhancing the defect image by adopting 5 data enhancement modes of translation, brightness change, noise addition, rotation and mirror image;

(7) And (3) carrying out the following steps on the processed underground drainage pipeline defect image data set according to the weight ratio of 8:1:1, dividing the training set, the verification set and the test set in proportion;

(8) Inputting the defect image data set of the underground drainage pipeline into an improved YOLOX algorithm for defect detection and identification;

(9) And testing the network obtained by training, counting the test result, and analyzing the accuracy rate, recall rate, average accuracy and average accuracy mean value of the algorithm.

Images